# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from __future__ import print_function import framework from framework import Program, default_main_program, default_startup_program, Parameter, Variable import optimizer from layer_helper import LayerHelper from distributed_spliter import * import math from . import core import debuger class VarBlock: def __init__(self, varname, offset, size): self.varname = varname # NOTE: real offset is offset * size self.offset = offset self.size = size def __str__(self): return "%s:%d:%d" % (self.varname, self.offset, self.size) class UnionFind(object): """ Union-find data struct. Union-find is a data struct that keeps track of a set of elements partitioned into a number of disjoint (non-overlapping) subsets. Reference: https://en.wikipedia.org/wiki/Disjoint-set_data_structure Args: elements(list): The initialize element list. """ def __init__(self, elementes=None): self._parents = [] # index -> parent index self._index = {} # element -> index self._curr_idx = 0 if not elementes: elementes = [] for ele in elementes: self._parents.append(self._curr_idx) self._index.update({ele: self._curr_idx}) self._curr_idx += 1 def find(self, x): # Find the root index of given element x, # execute the path compress while findind the root index if not x in self._index: return -1 idx = self._index[x] while idx != self._parents[idx]: t = self._parents[idx] self._parents[idx] = self._parents[t] idx = t return idx def union(self, x, y): # Union two given element x_root = self.find(x) y_root = self.find(y) if x_root == y_root: return self._parents[x_root] = y_root def is_connected(self, x, y): # If two given elements have the same root index, # then they are connected. return self.find(x) == self.find(y) def same_or_split_var(p_name, var_name): return p_name == var_name or p_name.startswith(var_name + ".block") def split_dense_variable(var_list, pserver_count, min_block_size=1024, max_block_size=1048576): """ We may need to split dense tensor to one or more blocks and put them equally onto parameter server. One block is a sub-tensor aligned by dim[0] of the tensor. We need to have a minimal block size so that the calculations in the parameter server side can gain better performance. By default minimum block size is 1024. The max block size is used to prevent very large blocks that may cause send error. """ blocks = [] for var in var_list: split_count = pserver_count var_numel = reduce(lambda x, y: x * y, var.shape) max_pserver_count = int(math.floor(var_numel / float(min_block_size))) if max_pserver_count == 0: max_pserver_count = 1 if max_pserver_count < pserver_count: split_count = max_pserver_count block_size = int(math.ceil(var_numel / float(split_count))) if len(var.shape) >= 2: # align by dim1(width) dim1 = reduce(lambda x, y: x * y, var.shape[1:]) remains = block_size % dim1 if remains != 0: block_size += dim1 - remains # update split_count after aligning split_count = int(math.ceil(var_numel / float(block_size))) for block_id in xrange(split_count): curr_block_size = min(block_size, var_numel - ( (block_id) * block_size)) block = VarBlock(var.name, block_id, curr_block_size) blocks.append(str(block)) return blocks class DistributeTranspiler: def transpile(self, optimize_ops, params_grads, trainer_id, program=None, pservers="127.0.0.1:6174", trainers=1, split_method=round_robin): """ Transpile the program to distributed data-parallelism programs. The main_program will be transformed to use a remote parameter server to do parameter optimization. And the optimization graph will be put into a parameter server program. Use different methods to split trainable variables to different parameter servers. Steps to transpile trainer: 1. split variable to multiple blocks, aligned by product(dim[1:]) (width). 2. rename splited grad variables to add trainer_id suffix ".trainer_%d". 3. modify trainer program add split_op to each grad variable. 4. append send_op to send splited variables to server and fetch params(splited blocks or origin param) from server. 5. append concat_op to merge splited blocks to update local weights. Steps to transpile pserver: 1. create new program for parameter server. 2. create params and grad variables that assigned to current server instance. 3. create a sub-block in the server side program 4. append ops that should run on current server instance. 5. add listen_and_serv op :param optimize_ops: op list of optimization, should be the return value of Optimizer.minimize :type optimize_ops: list :param params_grads: list of tuple(weight, gradient) :type params_grads: list :param trainer_id: one unique id for each trainer in a job. :type trainer_id: int :param program: program to transpile, default is default_main_program :type program: Program :param pservers: parameter server endpoints like "m1:6174,m2:6174" :type pservers: string :param trainers: total number of workers/trainers in the job :type trainers: int :param split_method: A function to determin how to split variables to different servers equally. :type split_method: function """ assert (callable(split_method)) if program is None: program = default_main_program() self.program = program self.trainers = trainers self.optimize_ops = optimize_ops # TODO(typhoonzero): currently trainer_id is fetched from cluster system # like Kubernetes, we should port this to use etcd later when developing # fluid distributed training with fault-tolerance. self.trainer_id = trainer_id pserver_endpoints = pservers.split(",") # step1 param_list = [pg[0] for pg in params_grads] grad_list = [pg[1] for pg in params_grads] grad_blocks = split_dense_variable(grad_list, len(pserver_endpoints)) param_blocks = split_dense_variable(param_list, len(pserver_endpoints)) # step2 grad_var_mapping = self._append_split_op(program, grad_blocks) # step3 send_inputs = [] send_outputs = [] for b in grad_blocks: # append by order varname, block_id, _ = b.split(":") send_inputs.append(grad_var_mapping[varname][int(block_id)]) param_var_mapping = self._create_vars_from_blocklist(program, param_blocks) for b in param_blocks: varname, block_id, _ = b.split(":") send_outputs.append(param_var_mapping[varname][int(block_id)]) # let send_op know which endpoint to send which var to, eplist has the same # order as send_inputs. eplist = split_method(send_inputs, pserver_endpoints) # create mapping of endpoint -> split var to create pserver side program self.param_grad_ep_mapping = dict() for i, ep in enumerate(eplist): param = send_outputs[i] grad = send_inputs[i] if not self.param_grad_ep_mapping.has_key(ep): self.param_grad_ep_mapping[ep] = {"params": [], "grads": []} self.param_grad_ep_mapping[ep]["params"].append(param) self.param_grad_ep_mapping[ep]["grads"].append(grad) rpc_client_var = program.global_block().create_var( name="RPC_CLIENT_VAR", persistable=True, type=core.VarDesc.VarType.RAW) # create send_op program.global_block().append_op( type="send", inputs={"X": send_inputs}, outputs={"Out": send_outputs, "RPCClient": rpc_client_var}, attrs={"endpoints": pserver_endpoints, "epmap": eplist}) # step4 for varname, splited_var in param_var_mapping.iteritems(): if len(splited_var) <= 1: continue orig_param = program.global_block().vars[varname] program.global_block().append_op( type="concat", inputs={"X": splited_var}, outputs={"Out": [orig_param]}, attrs={"axis": 0}) def get_trainer_program(self): # remove optimize ops and add a send op to main_program self.program.global_block().delete_ops(self.optimize_ops) # FIXME(typhoonzero): serialize once will fix error occurs when clone. self.program.__str__() return self.program def get_pserver_program(self, endpoint): """ Get pserver side program using the endpoint. NOTE: assume blocks of the same variable is not distributed on the same pserver, only change param/grad varnames for trainers to fetch. """ # step1 pserver_program = Program() # step2 recv_inputs = [] for v in self.param_grad_ep_mapping[endpoint]["params"]: self._clone_var(pserver_program.global_block(), v) for v in self.param_grad_ep_mapping[endpoint]["grads"]: # create vars for each trainer in global scope, so # we don't need to create them when grad arrives. # change client side var name to origin name by # removing ".trainer_%d" suffix suff_idx = v.name.find(".trainer_") if suff_idx >= 0: orig_var_name = v.name[:suff_idx] else: orig_var_name = v.name single_trainer_var = pserver_program.global_block().create_var( name=orig_var_name, persistable=True, type=v.type, dtype=v.dtype, shape=v.shape) if self.trainers > 1: for trainer_id in xrange(self.trainers): var = pserver_program.global_block().create_var( name="%s.trainer_%d" % (orig_var_name, trainer_id), persistable=False, type=v.type, dtype=v.dtype, shape=v.shape) recv_inputs.append(var) else: recv_inputs.append(single_trainer_var) # step3 optimize_block = pserver_program.create_block(0) # step 4 # Create a union-find data struct from optimize ops, # If two ops are connected, we could add these two ops # into one set. ufind = self._create_ufind(self.optimize_ops) # step 4.2 # Iterate through the ops and append optimize op which # located on current pserver opt_op_on_pserver = [] for _, op in enumerate(self.optimize_ops): if self._is_opt_op(op) and self._is_opt_op_on_pserver(endpoint, op): opt_op_on_pserver.append(op) # step 4.3 # Iterate through the ops, and if an op and the optimize ops # which located on current pserver are in one set, then # append it into the sub program. # We try to put optimization program run parallelly, assume # optimization program always looks like: # # prevop -> prevop -> opt op -> following op -> following op; -> # prevop -> prevop -> opt op -> following op -> following op; -> # global op -> global op # # we put operators that can run parallelly to many program blocks. # in above example, we seperate ops by the ";". Global ops must run # after all the optimize ops finished. global_ops = [] # HACK: optimization global ops only used to scale beta1 and beta2 # replace it with dependency engine. for op in self.optimize_ops: if op.type == "scale": for in_name in op.input_arg_names: if in_name.startswith("beta1_pow_acc") or\ in_name.startswith("beta2_pow_acc"): global_ops.append(op) def __append_optimize_op__(op, block): if self._is_opt_op(op): self._append_pserver_ops(block, op, endpoint, default_main_program()) else: self._append_pserver_non_opt_ops(block, op) append_block = optimize_block # append lr decay ops to the child block if exits lr_ops = self._get_lr_ops() if len(lr_ops) > 0: for _, op in enumerate(lr_ops): self._append_pserver_non_opt_ops(append_block, op) append_block = pserver_program.create_block(append_block.idx) # append op to the current block per_opt_block = append_block for _, opt_op in enumerate(opt_op_on_pserver): for _, op in enumerate(self.optimize_ops): # optimizer is connected to itself if ufind.is_connected(op, opt_op) and \ op not in global_ops: __append_optimize_op__(op, per_opt_block) per_opt_block = pserver_program.create_block(append_block.idx) # append global ops for glb_op in global_ops: __append_optimize_op__(glb_op, per_opt_block) # NOT USED: single block version: # # for _, op in enumerate(self.optimize_ops): # for _, opt_op in enumerate(opt_op_on_pserver): # if ufind.is_connected(op, opt_op): # __append_optimize_op__(glb_op, optimize_block) # break # step5 append the listen_and_serv op pserver_program.global_block().append_op( type="listen_and_serv", inputs={'X': recv_inputs}, outputs={}, attrs={ "OptimizeBlock": optimize_block, "endpoint": endpoint, "Fanin": self.trainers }) pserver_program.sync_with_cpp() return pserver_program def get_startup_program(self, endpoint, pserver_program): """ Get startup program for current parameter server. Modify operator input variables if there are variables that were split to several blocks. """ s_prog = Program() orig_s_prog = framework.default_startup_program() params = self.param_grad_ep_mapping[endpoint]["params"] def _get_splited_name_and_shape(varname): for idx, splited_param in enumerate(params): pname = splited_param.name if same_or_split_var(pname, varname) and varname != pname: return pname, splited_param.shape return "", [] # 1. create vars in pserver program to startup program pserver_vars = pserver_program.global_block().vars created_var_map = dict() for _, var in pserver_vars.iteritems(): tmpvar = s_prog.global_block().clone_variable(var) created_var_map[var.name] = tmpvar # 2. rename op outputs for op in orig_s_prog.global_block().ops: new_inputs = dict() new_outputs = dict() # do not append startup op if var is not on this pserver op_on_pserver = False for key in op.output_names: newname, _ = _get_splited_name_and_shape(op.output(key)[0]) if newname: op_on_pserver = True new_outputs[key] = created_var_map[newname] elif op.output(key)[0] in pserver_vars: op_on_pserver = True new_outputs[key] = pserver_vars[op.output(key)[0]] # most startup program ops have no inputs new_inputs = self._get_input_map_from_op(pserver_vars, op) if op_on_pserver: if op.type in [ "gaussian_random", "fill_constant", "uniform_random" ]: op.attrs["shape"] = new_outputs["Out"].shape s_prog.global_block().append_op( type=op.type, inputs=new_inputs, outputs=new_outputs, attrs=op.attrs) return s_prog # ====================== private transpiler functions ===================== def _create_vars_from_blocklist(self, program, block_list, add_trainer_suffix=False): """ NOTE: only grads need to be named for different trainers, use add_trainer_suffix to rename the grad vars. """ block_map = dict() var_mapping = dict() for block_str in block_list: varname, offset, size = block_str.split(":") if not block_map.has_key(varname): block_map[varname] = [] block_map[varname].append((long(offset), long(size))) for varname, splited in block_map.iteritems(): orig_var = program.global_block().var(varname) if len(splited) == 1: if add_trainer_suffix: new_var_name = "%s.trainer_%d" % \ (orig_var.name, self.trainer_id) program.global_block().rename_var(varname, new_var_name) var_mapping[varname] = \ [program.global_block().var(new_var_name)] else: var_mapping[varname] = \ [program.global_block().var(orig_var.name)] continue var_mapping[varname] = [] orig_shape = orig_var.shape orig_dim1_flatten = 1 if len(orig_shape) >= 2: orig_dim1_flatten = reduce(lambda x, y: x * y, orig_shape[1:]) for i, block in enumerate(splited): size = block[1] rows = size / orig_dim1_flatten splited_shape = [rows] if len(orig_shape) >= 2: splited_shape.extend(orig_shape[1:]) new_var_name = "" if add_trainer_suffix: new_var_name = "%s.block%d.trainer_%d" % \ (varname, i, self.trainer_id) else: new_var_name = "%s.block%d" % \ (varname, i) var = program.global_block().create_var( name=new_var_name, persistable=False, dtype=orig_var.dtype, type=orig_var.type, shape=splited_shape) # flattend splited var var_mapping[varname].append(var) program.global_block().sync_with_cpp() return var_mapping def _clone_var(self, block, var): assert isinstance(var, Variable) return block.create_var( name=var.name, shape=var.shape, dtype=var.dtype, type=var.type, lod_level=var.lod_level, persistable=True) def _append_split_op(self, program, gradblocks): # Split variables that need to be split and append respective ops add_suffix = False if self.trainers > 1: add_suffix = True var_mapping = self._create_vars_from_blocklist( program, gradblocks, add_trainer_suffix=add_suffix) for varname, splited_vars in var_mapping.iteritems(): # variable that don't need to split have empty splited_vars if len(splited_vars) <= 1: continue orig_var = program.global_block().vars[varname] if orig_var.type == core.VarDesc.VarType.SELECTED_ROWS: height_sections = [] for v in splited_vars: height_sections.append(v.shape[0]) program.global_block().append_op( type="split_selected_rows", inputs={"X": orig_var}, outputs={"Out": splited_vars}, attrs={"height_sections": height_sections}) elif orig_var.type == core.VarDesc.VarType.LOD_TENSOR: sections = [] for v in splited_vars: sections.append(v.shape[0]) program.global_block().append_op( type="split", inputs={"X": orig_var}, outputs={"Out": splited_vars}, attrs={"sections": sections} # assume split evenly ) else: AssertionError("Variable type should be in set " "[LOD_TENSOR, SELECTED_ROWS]") return var_mapping def _get_optimizer_input_shape(self, op_type, varkey, orig_shape, param_shape): """ Returns the shape for optimizer inputs that need to be reshaped when Param and Grad is split to multiple servers. """ # HACK(typhoonzero): Should use functions of corresponding optimizer in # optimizer.py to get the shape, do not bind this in the transpiler. if op_type == "adam": if varkey in ["Moment1", "Moment2"]: return param_shape elif op_type == "adagrad": if varkey == "Moment": return param_shape elif op_type == "adamax": if varkey in ["Moment", "InfNorm"]: return param_shape elif op_type == "momentum": if varkey == "Velocity": return param_shape elif op_type == "": if varkey == "Moment": return param_shape elif op_type == "sgd": pass return orig_shape def _orig_varname(self, varname): suff_idx = varname.find(".trainer_") orig_var_name = "" if suff_idx >= 0: orig_var_name = varname[:suff_idx] else: orig_var_name = varname return orig_var_name def _append_pserver_ops(self, optimize_block, opt_op, endpoint, origin_program): program = optimize_block.program pserver_block = program.global_block() new_inputs = dict() # update param/grad shape first, then other inputs like # moment can use the updated shape for key in opt_op.input_names: if key == "Grad": grad_block = None for g in self.param_grad_ep_mapping[endpoint]["grads"]: if same_or_split_var( self._orig_varname(g.name), self._orig_varname(opt_op.input(key)[0])): grad_block = g break if not grad_block: # do not append this op if current endpoint # is not dealing with this grad block return merged_var = \ pserver_block.vars[self._orig_varname(grad_block.name)] if self.trainers > 1: vars2merge = [] for i in xrange(self.trainers): per_trainer_name = "%s.trainer_%d" % \ (self._orig_varname(grad_block.name), i) vars2merge.append(pserver_block.vars[per_trainer_name]) optimize_block.append_op( type="sum", inputs={"X": vars2merge}, outputs={"Out": merged_var}) if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS: optimize_block.append_op( type="scale", inputs={"X": merged_var}, outputs={"Out": merged_var}, attrs={"scale": 1.0 / float(self.trainers)}) new_inputs[key] = merged_var elif key == "Param": # param is already created on global program param_block = None for p in self.param_grad_ep_mapping[endpoint]["params"]: if same_or_split_var(p.name, opt_op.input(key)[0]): param_block = p break if not param_block: return tmpvar = pserver_block.create_var( name=param_block.name, persistable=True, dtype=param_block.dtype, shape=param_block.shape) new_inputs[key] = tmpvar elif key == "LearningRate": # leraning rate variable has already be created by non-optimize op, # don't create it once again. lr_varname = opt_op.input(key)[0] if pserver_block.vars.has_key(lr_varname): new_inputs[key] = pserver_block.vars[opt_op.input(key)[0]] else: origin_var = origin_program.global_block().vars[lr_varname] tmpvar = pserver_block.create_var( name=origin_var.name, persistable=origin_var.persistable, dtype=origin_var.dtype, shape=origin_var.shape) new_inputs[key] = tmpvar for key in opt_op.input_names: new_shape = None if key in ["Param", "Grad", "LearningRate"]: continue var = self.program.global_block().vars[opt_op.input(key)[0]] # update accumulator variable shape param_shape = new_inputs["Param"].shape new_shape = self._get_optimizer_input_shape(opt_op.type, key, var.shape, param_shape) tmpvar = pserver_block.create_var( name=var.name, persistable=var.persistable, dtype=var.dtype, shape=new_shape) new_inputs[key] = tmpvar # change output's ParamOut variable outputs = self._get_output_map_from_op(self.program.global_block().vars, opt_op) outputs["ParamOut"] = new_inputs["Param"] optimize_block.append_op( type=opt_op.type, inputs=new_inputs, outputs=outputs, attrs=opt_op.attrs) def _append_pserver_non_opt_ops(self, optimize_block, opt_op): program = optimize_block.program # Append the ops for parameters that do not need to be optimized/updated inputs = self._get_input_map_from_op(self.program.global_block().vars, opt_op) for varlist in inputs.itervalues(): if not isinstance(varlist, list): varlist = [varlist] for var in varlist: if not program.global_block().vars.has_key(var.name): program.global_block().create_var( name=var.name, persistable=var.persistable, dtype=var.dtype, shape=var.shape) outputs = self._get_output_map_from_op(self.program.global_block().vars, opt_op) for varlist in outputs.itervalues(): if not isinstance(varlist, list): varlist = [varlist] for var in varlist: program.global_block().clone_variable(var) optimize_block.append_op( type=opt_op.type, inputs=inputs, outputs=outputs, attrs=opt_op.attrs) def _is_op_connected(self, op1, op2): # If one op's input is another op's output or # one op's output is another op's input, we say # the two operator is connected. def _append_inname_remove_beta(varname_list): op_input_names = [] for in_name in varname_list: # HACK: remove beta1 and beta2 to avoid let all # ops connected. if in_name.startswith("beta2_pow_acc") or \ in_name.startswith("beta1_pow_acc"): continue else: op_input_names.append(in_name) return op_input_names op1_input_names = _append_inname_remove_beta(op1.desc.input_arg_names()) op1_output_names = op1.desc.output_arg_names() op2_input_names = _append_inname_remove_beta(op2.desc.input_arg_names()) op2_output_names = op2.desc.output_arg_names() if set(op1_output_names) & set(op2_input_names) or \ set(op1_input_names) & set(op2_output_names): return True return False def _create_ufind(self, optimize_ops): # Create a unit find data struct by optimize ops ufind = UnionFind(optimize_ops) for i in xrange(len(optimize_ops)): for j in xrange(i, len(optimize_ops)): op1 = optimize_ops[i] op2 = optimize_ops[j] if self._is_op_connected(op1, op2): ufind.union(op1, op2) return ufind def _is_opt_op(self, op): # NOTE: It's a HACK implement. # optimize op: SGDOptimize, MomentumOptimizer, AdamOptimizer and etc... if "Param" in op.input_names and \ "LearningRate" in op.input_names: return True return False def _is_opt_op_on_pserver(self, endpoint, op): param_names = [ p.name for p in self.param_grad_ep_mapping[endpoint]["params"] ] if op.input("Param")[0] in param_names: return True else: for n in param_names: param = op.input("Param")[0] if same_or_split_var(n, param) and n != param: return True return False return False def _get_input_map_from_op(self, varmap, op): iomap = dict() for key in op.input_names: vars = [] for varname in op.input(key): vars.append(varmap[varname]) if len(vars) == 1: iomap[key] = vars[0] else: iomap[key] = vars return iomap def _get_output_map_from_op(self, varmap, op): iomap = dict() for key in op.output_names: vars = [] for varname in op.output(key): vars.append(varmap[varname]) if len(vars) == 1: iomap[key] = vars[0] else: iomap[key] = vars return iomap def _get_lr_ops(self): lr_ops = [] # find learning rate variables by optimize op lr_vars = set() for op in self.optimize_ops: if self._is_opt_op(op): lr_vars.add(op.input("LearningRate")[0]) find_ops = [] # find ops which output is lr var block = self.program.global_block() for op in block.ops: if set(op.output_arg_names) & lr_vars: find_ops.append(op) # make a union find struct by the ops in default_main_program ufind = UnionFind(block.ops) for op1 in block.ops: for op2 in block.ops: # NOTE: we need to skip all optimize ops, since it is connected # with forward/backward ops and lr ops, we only need the lr ops. if op1 != op2 and self._is_op_connected(op1, op2) and \ not self._is_opt_op(op1) and not self._is_opt_op(op2): ufind.union(op1, op2) # find all ops which is related with lr var for op1 in block.ops: for op2 in find_ops: if ufind.is_connected(op1, op2): lr_ops.append(op1) return lr_ops